Bayesian Optimized Enhanced Ensemble Multi-Layer Perceptron for Rear-End Collision Prediction in IoV
Dilna Vijayan, Mohamed Saad, Ahmed M. Khedr · IEEE Access · 2024
The trend of creating and deploying intelligent transportation systems within the context of the Internet of Vehicles (IoV) concept is gaining traction and attracting interest from both academia and industry. Rear-end collisions, which constitute an immense amount of traffic accidents, are an essential concern in an effort to improve road safety in smart cities. The complexity entailed in tackling this problem, specifically concerning the prediction of rear-end crashes, demands the application of cutting-edge methodologies. Learning-based techniques have shown promise, but traditional methods are not very successful in offering solutions. Nevertheless, there are drawbacks to these techniques, especially with regard to feature extraction and prediction accuracy. This study presents the Bayesian Optimized Enhanced Ensemble Multi-Layer Perceptron, a novel method for predictive modeling in IoV safety for rear-end collision prediction. Our approach leverages the feature extraction capabilities of Convolutional Neural Networks (CNNs) to interpret complex sensor data from connected vehicles. In order to expedite experimentation, we utilize a meticulously selected subset of an extensive dataset that includes insightful records about vehicle behavior and collision incidents. This dataset undergoes preprocessing and normalization prior to one-dimensional (1D) CNN processing for feature extraction. Furthermore, our predictive modeling approach employs an ensemble of Multi-Layer Perceptrons (MLPs). In order to maximize model performance, hyper-parameters are optimized. We compare the effectiveness of our proposed model against alternative methods and show its superiority in predicting collision events through experimentation and rigorous evaluation. The results show that our suggested solution outperforms existing methods and considerably improves IoV safety systems. Specifically, the suggested model achieved an MSE of 0.00004403 and an RMSE of 0.006617, exhibiting a 30.03% improvement in RMSE over ANN, 27.51% over DRNN-LSTM, 24.71% over EMLP, 19.28% over BPNN, and 27.50% over A3C. These results highlight the model’s ability to reliably forecast rear-end crashes, which is critical for improving driver safety and reducing accident risks in traffic scenarios.